%0 = stablehlo.convert %arg0 : (tensor<1x2xf32>) -> tensor<1x2xi8>
    return %0 : tensor<1x2xi8>
  }
// CHECK: @uniform_quantize_0
  func.func private @uniform_quantize_1(%arg0: tensor<1x3xf32>, %arg1: tensor<1x1xf32>, %arg2: tensor<1x1xi8>) -> tensor<1x3xi8> {
    %0 = stablehlo.convert %arg0 : (tensor<1x3xf32>) -> tensor<1x3xi8>
    return %0 : tensor<1x3xi8>
  }

      %1 = stablehlo.uniform_quantize %arg0 : (tensor<1x2xf32>) -> tensor<1x2x!quant.uniform<i8:f32, 4.000000e+0>>

  %zps = "tf.Const"() { value = dense<3> : tensor<i32> } : () -> tensor<i32>

  // CHECK: %[[input:.*]] = "tf.ConcatV2"(%arg0, %arg1, %[[VAL:.*]]) : (tensor<3x3xi8>, tensor<3x3xi8>, tensor<i64>) -> tensor<6x3xi8>
  // CHECK: %[[input_qint:.*]] = "tf.Cast"(%[[input]]) <{Truncate = false}> : (tensor<6x3xi8>) -> tensor<6x3x!tf_type.qint8>
  // CHECK: %[[output:.*]] = "tf.UniformDequantize"(%[[input_qint]]
  // CHECK: return %[[output]] : tensor<6x3xf32>

  // CHECK-LABEL: func.func private @merge_relu_fusion
  func.func private @merge_relu_fusion(%arg0: tensor<1x4xf32>) -> tensor<1x3xf32> {
    %0 = stablehlo.constant() {value = dense<127> : tensor<4x3xi8>} : () -> tensor<4x3x!quant.uniform<i8<-127:127>:f32:1, {5.000000e-03,5.000000e-03,5.000000e-03}>>
    %1 = stablehlo.uniform_quantize %arg0 : (tensor<1x4xf32>) -> tensor<1x4x!quant.uniform<i8:f32, 6.000000e-03:-128>>

  return %2 : tensor<?x2x2x1xf32>
}

// -----

// CHECK-LABEL: func @dot_general_add_add
func.func @dot_general_add_add(
    %lhs: tensor<?x?x6xi8>, %rhs: tensor<6x8xi8>,
    %zp_offset: tensor<8xi32>, %bias: tensor<8xi32>
  ) -> tensor<?x?x8xi32> {
  // CHECK: %[[dot:.*]] = "mhlo.dot_general"
  // CHECK: %[[combined:.*]] = chlo.broadcast_add %[[zp_offset:.*]], %[[bias:.*]]

}

func.func @RemoveTrivialSubInt8(%arg0: tensor<2x2xi8>) -> tensor<2x2xi8> {
  %cst = arith.constant dense<0> : tensor<2x2xi8>
  %0 = "tf.Sub"(%arg0, %cst) : (tensor<2x2xi8>, tensor<2x2xi8>) -> tensor<2x2xi8>
  func.return %0 : tensor<2x2xi8>

  // CHECK-LABEL: RemoveTrivialSubInt8
  // CHECK-NEXT: return %arg0 : tensor<2x2xi8>
}

func.func @RemoveTrivialMul(%arg0: tensor<2x2xf32>) -> tensor<2x2xf32> {

  // CHECK-SAME: %[[ARG0:.*]]: tensor<1x2xf32>
  // CHECK-SAME: %[[ARG1:.*]]: tensor<2x3xf32>
  // CHECK-SAME: %[[ARG2:.*]]: tensor<1x3xi1>
  // CHECK-SAME: %[[ARG3:.*]]: tensor<1x3xf32>
  func.func private @composite_and_select(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>, %arg2: tensor<1x3xi1>, %arg3: tensor<1x3xf32>) -> tensor<1x3xf32> {

  // CHECK: %[[WHILE_RESULT:.*]]:2 = mhlo.while(%[[ARG0:.*]] = %[[INT0]], %[[ARG1:.*]] = %[[INT1]])
  // CHECK-SAME: tensor<2x2xi8>, tensor<2x2xi8>

  // CHECK: cond

  // CHECK: do
  // CHECK: mhlo.return %[[ARG0]], %[[ARG1]] : tensor<2x?xi8>, tensor<?x2xi8>

  %2:2 = "tf.WhileRegion"(%0, %1) ({
  ^bb0(%carg0: tensor<2x2x!tf_type.qint8>, %carg1: tensor<2x2x!tf_type.qint8>):

// PerTensor-SAME: f = @quantized_matmul_fn_0}> : (tensor<2x12xf32>, tensor<12x2xi8>, tensor<f32>, tensor<i32>) -> tensor<*xf32>
// PerTensor: return %[[out]]

// PerChannel-LABEL: func @matmul
// PerChannel-DAG: %[[q_w:.*]] = "tf.Const"() <{value = dense<0> : tensor<12x2xi8>}> : () -> tensor<12x2xi8>
// PerChannel-DAG: %[[scale:.*]] = "tf.Const"() <{value = dense<3.93700805E-9> : tensor<f32>}> : () -> tensor<f32>

  // CHECK: %[[BROADCAST:.*]] = "mhlo.broadcast_in_dim"(%arg0) <{broadcast_dimensions = dense<1> : tensor<1xi64>}> : (tensor<4xui8>) -> tensor<2x4xui8>
  // CHECK:  mhlo.shift_right_logical %[[BROADCAST]], %arg1 : tensor<2x4xui8>
  %0 = "tf.RightShift"(%arg0, %arg1) : (tensor<4xui8>, tensor<2x4xui8>) -> tensor<2x4xui8>
  func.return %0 : tensor<2x4xui8>
}